Owain Lewis

@ryancarson I don’t understand the hype. Most of the use cases seem like trivial problems you could more easily do yourself. When has scheduling a meeting ever needed an AI?

Important reminder to be careful online. The link here “upgrade Claude Code” is a Google sponsored ad that installs malware on your machine.

@Al_Grigor @hnshah $ Would you like me to delete the production database? Y/N $ Yes please go ahead

Claude Code wiped our production database with a Terraform command. It took down the DataTalksClub course platform and 2.5 years of submissions: homework, projects, and leaderboards. Automated snapshots were gone too. In the newsletter, I wrote the full timeline + what I changed so this doesn't happen again. If you use Terraform (or let agents touch infra), this is a good story for you to read. alexeyondata.substack.com/p/how-i-droppe…

@Al_Grigor @hnshah Seems like a process problem not a tool problem. Terraform is equally capable of deleting resources if you let it.

What's up with Vertex AI quota? I need it for client work and get instantly auto denied so basically can't use it and have no way of finding out why. How do you get quota? @googlecloud

The best code isn't written. It's rewritten. No one gets it perfect on the first pass - not you, not AI. AI gives you a fast first draft. The craft is what happens next: review it, simplify it, make it fit your codebase. Treat every AI output as a starting point, not a finished product.

@byzkhan That's fair. I prefer Claude for work but I'd be equally happy with most of the agent tools right now.

@owainlewis tools used within the workflow matter though

Claude Code or Cursor? OpenCode or Codex? Wrong question. The tool doesn't matter. The workflow does. And the workflow is the same one engineers have used for decades: Design → Plan → Build → Review → Commit Just faster now.

Most people take AI output and either accept it or complain. The unlock: treat the first output as a draft, not a final answer. "What edge cases did we miss?" "What would a senior engineer change?" The best results come from iteration, not prompting.

AI YouTube right now: "I vibe coded this with AI. I don’t have a clue how it works. Selling it to clients for $10k."

The skill that makes you better with AI: Thinking in standard operating procedures (SOPs). 1. Collect examples of what "good" looks like 2. Turn them into an SOP The clearer your system, the better any model performs. AI doesn't need clever prompts. It needs clear instructions.

Good software engineering isn't picking the most complex solution. Or the one everyone's talking about online. My mantra: “Build boring software” Don’t be afraid of simple solutions. They’re often the best.

@nicoloboschi Great question. I've seen it happen in work because there's so much content about RAG and vector DBs, people just assume it's "the (only) way".

@owainlewis yeah but if you have 10 files only, how did you ever end up thinking you needed a retrieval system?

RAG doesn't mean vector databases. I've seen engineers spin up complex database infrastructure to search 10 small documents when they could just... load the file. For small doc sets, picking the right file and reading it whole beats chunking and retrieval every time. More reliable, less code, no infra. Simple gets ignored because it doesn't feel like "real" AI. But simple solutions are usually the best.

Two types of thinking about AI: Type 1: "How do I do my work faster?" Type 2: "How do I build a system that does better work than me, at a scale I couldn't match?" Type 1 saves you 10 minutes here and there. You're still the bottleneck. Type 2 creates something superhuman. One is optimisation. The other is a completely different game. Type 2 requires letting go of your ego (“only I can do this work”). The opportunity isn't saving time. It's building systems that do work humans simply couldn't do.

Your job as an engineer is changing. You're no longer the one who types the code. You're the one who builds the systems that do. Some engineers have already figured this out. They stopped competing where they can't win. They handed over the mechanical work and kept what machines can't do: the critical thinking, the taste, the judgment. Your new job is teaching machines to build better than you could alone. Your mind is the bottleneck now. Think bigger.

The engineers pulling ahead aren't better coders. They just stopped doing work machines can do faster. They let go of the syntax. The boilerplate. The implementation details. They held onto the decisions. The architecture. The taste. That's the difference now. Not how fast you type. But what you refuse to type at all.

Your job isn't to write code anymore. It's to teach systems to write code for you.

Four ways bad context breaks your agent: 1. Context poisoning: a hallucination enters context and corrupts future reasoning 2. Context distraction: too much context overwhelms the model 3. Context confusion: irrelevant info influences responses 4. Context clash: different parts of context contradict each other If your agent works at first then drifts later, one of these is usually why. All of these cause the same problem: agent says/does the wrong thing. This is why people talk about context engineering.

Four ways to manage context when building AI agents: Write: save state to files, not context Select: load data just-in-time, not upfront Compress: summarise or prune when context grows Isolate: split work across multiple agents (each with isolated context) The goal = only include what matters for the next step.